pantsbuild / pants / 38595183949

93%

push

github

Resolve Init and Conftest Owners from a Run-wide Index (#23699)

For every Python file, init and conftest inference ran `find_owners` and
then `resolve_targets` on each ancestor `__init__.py` / `conftest.py`.
Every file in a package repeated the same owner lookups.

This adds `python_source_owners_by_file`, which maps each file owned by
a target with a `PythonSourceField` to the owners' addresses and their
normalized resolves. It's built once per run from `AllPythonTargets`,
which import inference already computes for the module mapping. A target
that owns a file is declared in the file's directory or an ancestor. So
the index holds exactly the owners `find_owners` would return after the
existing `PythonSourceField` filter.

- `__init__` inference looks up each ancestor file in the index, then
filters by resolve as before.
- Conftest inference uses the index for files that are in it. A
`conftest.py` missing from the index still goes through `find_owners`
with `owners_not_found_behavior=error`, so an unowned conftest still
raises the same error.

Trade-off: when import inference is off but init/conftest inference is
on, a run that touches only a few files now computes `AllPythonTargets`,
which needs every target in the repo. It didn't before. With import
inference on (the default), that's already computed. I found very few
examples on github that would be impacted by that, and on those repos it
doesn't make a difference.

This is between an ~8% and a ~15% reduction in wall clock time over a
few different repositories with both a large number of targets and a
large number of `__init__.py` and `conftest.py` files.

52 of 52 new or added lines in 2 files covered. (100.0%)

96314 of 103628 relevant lines covered (92.94%)

3.66 hits per line